Nathanael,
BCP 54 only applies to the IETF stream. This document applies to all RFC
streams.
Regards/Ngā mihi
Brian Carpenter
On 06-Jun-26 17:35, Nathanael Ritz wrote:
Hi again,
Honestly, I think the friction here is that BCP 54 already covers much of what
authors are expected to affirm implicitly. In particular, the expectation that “IETF
participants use **their best engineering judgment** to find the best solution
[…]" excludes poor-faith submission of LLM synthesized output by construct, as
LLMs are incapable of synthesizing genuine *judgement*.
If we are asked to look at someone's contribution by email, GitHub PR or someone's -00 or or even
the -15 revision of someone's personal I-D, we are already expecting those submissions to be
filtered by the authors' "best engineering judgement" in according to BCP 54. So a
submission where "a substantial part of the document was created by AI"
that risks misleading the viewer would already violate the relevant provisions
in the existing Guidelines for Conduct,
making the AI disclosure requirement redundant.
Cheers,
Nathanael
On Fri, 5 Jun 2026 at 21:14, Brian E Carpenter <[email protected]
<mailto:[email protected]>> wrote:
On 06-Jun-26 14:06, John R Levine wrote:
>> Again, I'm not proposing this for copyright/legal reasons - that's
>> not RSWG business anyway - I'm talking about what RFC readers might
expeact
>> and care about.
>
> For about the fourth time, I think it is fine to ask, but if you're going
> to make it a requirement there's a lot more to do that I don't think we
> should do, at least not any time soon.
>
>> I decided to ask an expert, and here's what it said:
>> https://chatgpt.com/c/6a2379de-8228-83ec-8695-4c63d9fd08dd
<https://chatgpt.com/c/6a2379de-8228-83ec-8695-4c63d9fd08dd>
>
> It said I had to log in. Nope. As I'm sure you know, it's tuned to be
> sycophantic and tell you that whatever you say is brilliant.
Sorry about that - My question was "do people reading text want to know if it
was generated by a LLM?"
AI content is below my signature.
Brian
The evidence so far suggests a somewhat paradoxical answer:
Yes, most people say they want to know.
Surveys consistently find large majorities favor disclosure when AI has been
used. For example, one 2026 international survey found 86% of respondents wanted
AI-generated content to be disclosed, and journalism studies have reported similar
numbers above 80%.
[https://tribune.net.ph/2026/04/28/transparency-always-wins-a-new-global-study-finds-86-of-consumers-want-ai-generated-content-disclosed?utm_source=chatgpt.com
<https://tribune.net.ph/2026/04/28/transparency-always-wins-a-new-global-study-finds-86-of-consumers-want-ai-generated-content-disclosed?utm_source=chatgpt.com>]
But disclosure often makes people trust the text less.
Several experiments have found that labeling content as "AI-generated"
reduces perceived accuracy, authenticity, and willingness to share it—even when the
content itself is unchanged, true, or even human-written. And readers frequently cannot
tell the difference on their own.
In blind tests, people often rate AI-generated content as equal to or better than
human-written content. Some studies found a preference for AI-generated answers when
the source was hidden, with the preference declining once participants were told the
content came from AI.
[https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com
<https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com>]
So there are really two different questions:
Do readers want transparency?
Generally yes. Many people view disclosure as an ethical issue and want the
information available. [https://arxiv.org/abs/2503.16458?utm_source=chatgpt.com
<https://arxiv.org/abs/2503.16458?utm_source=chatgpt.com>]
Does knowing improve their evaluation of the text?
Not necessarily. In many studies, the label changes perceptions more than the
content itself does.
[https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com
<https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com>]
For a community such as the IETF, there is a further distinction. Readers
might not care whether an author used an LLM as a spelling checker, grammar
aid, or drafting assistant. They may care much more about whether the technical
content, protocol design, security analysis, and normative requirements were
actually reviewed and validated by the named authors. In that setting,
disclosure is often less about literary authorship and more about
accountability.
In short: people generally want to know, but they do not necessarily use that
information rationally when judging the quality of the text. The research literature
currently points quite strongly in that direction.
https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com
<https://www.researchgate.net/publication/384533200_People_are_skeptical_of_headlines_labeled_as_AI-generated_even_if_true_or_human-made_because_they_assume_full_AI_automation?utm_source=chatgpt.com>]
--
rswg mailing list -- [email protected] <mailto:[email protected]>
To unsubscribe send an email to [email protected]
<mailto:[email protected]>
--
rswg mailing list -- [email protected]
To unsubscribe send an email to [email protected]
